Heterogeneity in the Efficiency of Intrahousehold Resource
Allocation: Empirical Evidence and Implications for Investment
in Children
Manuela Angelucci∗ and Robert Garlick†
August 15, 2015
PRELIMINARY AND INCOMPLETE - PLEASE DO NOT CITE
Abstract
We present some of the first evidence of within-sample variation in the efficiency of intrahousehold resource allocation. In a sample of rural, low-income Mexican households, observed
consumption patterns are Pareto efficient for households with relatively old heads, but not households with relative young heads. This variation in efficiency has important welfare implications:
younger households invest less in children’s education and, in particular, their education investments are less sensitive to the receipt of cash transfers. The link between inefficient household
resource allocation and underprovision of public goods follows naturally from standard models but
we believe this has never before been documented empirically. Heterogeneity in efficiency by household head age may be due to cohort and/or lifecycle effects. The specific patterns of heterogeneity
we find are more consistent with cohort than lifecycle effects, which means that average efficiency
may decline over time in this population. From a policy perspective, these results highlight a
limitation of cash transfers relative to other forms of poverty alleviation policy in settings where
at least some households engage in inefficient consumption behaviour. Many papers have already
established that the distributional effects of development policy depend on whether households
behave as unitary actors; our results emphasize that it is also important to assess whether (some)
households behave as efficient collective units.
1
Introduction
Intrahousehold resource allocation is an important and widely-studied topic in economics. An extensive literature studies how households accumulate resources and allocate these resources between
∗ Assistant
† Assistant
Professor of Economics, University of Michigan; [email protected]
Professor of Economics, Duke University; [email protected]
1
members. Describing and modeling these decisions can provide important insights into households’
use of productive resources (Udry, 1996), investment in children’s human capital (Thomas, 1990), and
labor supply decisions (Rangel, 2006). In particular, many papers test whether households can be
accurately modeled as unitary actors with rational and consistent preferences and/or as cooperative
units whose consumption, investment, and production decisions are Pareto efficient. Most empirical
work rejects unitary models of household behavior but the evidence regarding cooperative models is
inconclusive.
We present a novel potential explanation for the range of results regarding the validity of cooperative
models of household behavior. We show that in a sample of rural Mexican households, the cooperative
model is strongly rejected for young households but not rejected for old households.1 Our tests are
based on empirical implications of static cooperative bargaining models that assume households always
reach Pareto efficient consumption allocations. Our result demonstrates that younger households
on average fail to achieve Pareto efficient consumption allocations and so “leave resources on the
table.” We believe that this is the first documentation of within-sample heterogeneity over observed
demographic characteristics in the efficiency of household resource allocations.2 This result is consistent
with both cohort and life-cycle factors. Households may be more likely to attain Pareto efficient
consumption allocations at older ages, perhaps because efficient bargaining is a learned skill or because
those households unable to achieve Pareto efficiency are more likely to dissolve. Alternatively, efficient
bargaining may be linked to some time-invariant characteristic that is more prevalent in older cohorts
in these rural Mexican states. We present suggestive evidence that our results are more consistent
with cohort factors, by comparing cross-sectional surveys at different points in time and by examining
cohort-level characteristics that vary minimally through the life cycle.
We make three contributions with this paper. First, and most important, our findings show that
the efficiency of household decision-making, and hence the validity of cooperative models of the household, can differ even when households face the same geographic, institutional and socio-economic
environment. This suggests that tests for “the validity of cooperative household models” may be misplaced; researchers should instead aim to identify which models best characterize different groups of
households. For example, the disparate results of prior empirical tests may reflect differences in the
age composition of samples across different studies. Standard approaches that test whether an entire
sample’s behaviour is consistent with a specific model do not take this into account. Second, we show
that inefficiency is more consistent with cohort than lifecycle explanations. This suggests that average
efficiency may fall over time, with important welfare implications. This may provide guidance for the
1 We define “young” and “old” households as households whose is head is respectively younger or older than the
sample median, 39. The results are robust to small changes in the value of this cutoff.
2 In related work, Angelucci (2008) shows that behavioral responses to cash transfers vary across rural Mexican households with different baseline characteristics. Farfán (2013) documents differences in resource-sharing and expenditure
for Mexican households depending on whether their members are co-resident in one household or split across multiple
locations. Hoel (2013) shows that asymmetric information prevents some but not all households from playing Pareto
efficient strategies in modified dictator games.
2
active research program into models of non-cooperative household behavior.
Third, we show inefficiency is associated with lower investment in children’s education and with
spending a smaller fraction of cash transfers on children’s education. This suggests that inefficiency is
associated with direct welfare costs for members of inefficient households and with lower effectiveness
of public anti-poverty policy. While cash transfers have been a popular policy measure in many
developing and developed countries, their effectiveness may be limited in environments where some
households bargain inefficiently and hence fail to take advantage of the potential welfare gains from
these transfers. We believe that we are the first to document an empirical relationship between
(in)efficient household bargaining and public goods provision. There is a clear theoretical prediction
that in inefficient households, members will contribute less to public goods such as children’s education.
However, no previous work has identified two different groups of households whose aggregate behaviour
respectively is and is not consistent with efficiency. This has made it impossible to test whether
inefficiency is indeed associated with lower public goods provision. Our approach relies on aggregating
households intro groups whose aggregate behaviour is more or less consistent with efficiency. We thus
cannot be sure that our results are not driven by unmodeled heterogeneity at the household level that
is correlated with (in)efficiency. But our main result is robust to controlling for the most obvious
measures of heterogeneity: differences in household resources and composition, and differences in local
market conditions.
Studies of household decision-making have considerably broader implications for understanding
group decision-making. Households may plausibly face more conducive conditions for achieving efficient outcomes than other groups. Household members have relatively extensive information on other
members’ behavior, face substantial costs of leaving the group, are more likely to have altruistic preferences toward each other, and may have relatively homogeneous preferences. If households are not
able to achieve efficient outcomes, the scope for other groups to do so may be even more limited.
We begin by outlining a theoretical and empirical framework in section 2. This framework closely
follows the canonical cooperative model of household behavior (Chiappori, 1992; Browning and Chiappori, 1998) and the alternative tests developed in Bourguignon, Browning, and Chiappori (2009).
We describe the setting and data in section 3. The data come from surveys conducted in rural Mexico
between 1997 and 2007 to evaluate the Progresa/Oportunidades program. We present our main results
in section 4, documenting the outcomes of efficiency tests for the full sample and the subsamples of
younger and older households. In section 5 we explore cohort and life-cycle explanations for the main
findings. We discuss relative education investments in more and less efficient households in section 6.
We conclude in section 7.
Our contribution builds off an extensive empirical and theoretical literature on intra-household
decision-making. A substantial body of empirical work fails to reject the null hypothesis that household
resource allocations are Pareto efficient (Akresh, 2005; Bobonis, 2009; Bourguignon, Browning, Chi-
3
appori, and Lechene, 1993; Browning, Bourguignon, Chiappori, and Lechene, 1994; Chiappori, Fortin,
and Lacroix, 2002; Rangel and Thomas, 2005). However, rejections of Pareto efficient consumption
decisions are not uncommon (Dercon and Krishnan, 2000; Djebbari, 2005) and several studies reject
Pareto efficient allocation of productive resources (Udry, 1996; Duflo and Udry, 2004). Several recent
papers explore possible reasons why Pareto efficiency is not achieved, including limited information
(Ashraf, 2009), heterogeneous time preferences (Schaner, 2013), and limited intertemporal commitment
(Mazzocco, 2007).
2
Theoretical and Empirical Framework: Testing the Cooperative Model
We lay out a simple model household consumption decisions to motivate the empirical work. This
model follows closely from Bourguignon, Browning, Chiappori, and Lechene (1993) and Browning,
Bourguignon, Chiappori, and Lechene (1994) and is fairly standard in the literature on household
bargaining. Households maximize the weighted sum of the utilities of members A and B
uA (q A , q B , Q; a) + µ(y, z, p) · uB (q A , q B , Q; a),
(1)
subject to a budget constraint
p·q ≤y
where q i is a vector of consumption of private goods for household member i, Q is a vector of consumption of public goods, q = (q A + q b , Q), p is a vector of prices, a is a vector of preference parameters,
and µ is the relative weight attached to member B’s utility. µ depends on a vector of distribution
factors z = (z 1 , . . . , z M ), which are typically interpreted as measures of bargaining power. We assume
that the utility functions uA and uB satsify standard conditions but do not impose any parametric
assumptions at this stage. The specification of the demand problem in equation 1 assumes that distribution factors influence equilibrium consumption only through the weight function µ, not through the
utility functions uA and uB . Substantively, this rules out candidate distribution factors from which
household members might derive direct utility.
The solution to this problem yields a reduced form household demand for good j
¯ µ(y, p, z), p, a) = d(y,
˜ z, p, a),
dj = d(y,
This demand function implies that the household budget shares for goods j ∈ {1, . . . , J} are given by
wj =
dj
= d(x, z, p, a)
x
4
(2)
where x =
PJ
j=1
wj is total expenditure.3
If the household behaves as a unitary actor, then intrahousehold resource allocations should not
depend on the bargining power of the individual members. Hence, conditional on total household
consumption, the distribution factors should not influence the distribution of consumption between
different goods and so
∂wj
∂z m
= 0 in equation 2 for all j ∈ {1, . . . , J} and m ∈ {1, . . . , M }. This provides
a testable restriction of the unitary model.
If the household’s allocation of resources is Pareto efficient, then the relative effect of changes in
different distribution factors on the budget share for each consumption good should be identical across
∂wj ∂wj
∂wk ∂wk
all goods: ∂z
m
∂z n = ∂z m ∂z n for all j, k ∈ {1, . . . , J} and m, n ∈ {1, . . . , M }. This is typically
called the joint proportionality condition. Intuitively, this condition is easiest to understand in terms
of a multi-stage budgeting process. Browning and Chiappori (1998) shows that the collective model
is formally equivalent to a model where households members first pool their incomes, then divide
incomes between members based on bargaining power, and members finally allocate their income
shares in accordance with their individual preferences. If a change in the distribution factor z m
increases member A’s bargaining power, her income share will increase in the division stage. She will
adjust her expenditure on each element of q based on her individual utility function uA (q A , q B , Q; a),
which does not depend on z. The household budget share vector w will then change, but this change
will be due only to the change in the income share received by A, not on which distribution factors
induced this change. Hence, the relative shifts in budget shares will be identical across distribution
factors, yielding the joint proportionality condition. This provides a testable restriction of the collective
or cooperative model. Note that this is a strictly weaker restriction on consumption behavior than that
implied by the unitary model. Hence, consumption behavior may be consistent with the cooperative
model and not the unitary model but not vice versa.
We impose some parametric structure on the budget share models in equation 2. While this
structure is restrictive, it is standard in the literature and nonparametric estimation of household
demand models is fairly rare (Cherchye, Rock, and Vermeulen, 2009). We begin by estimating Engel
curves derived from the almost ideal demand system (Deaton and Muellbauer, 1980):
j
wi,t
= αj ln(xi,t ) +
M
X
j
m
φjm zi,t
+ β j ai,t + ηi,t
+ ji,t
(3)
m=1
for each good j and for households i ∈ {1, . . . , n} in periods t ∈ {1, . . . , T }. xi,t represents total
1
M
consumption expenditure; zi,t
, . . . , zi,t
represent distribution factors; ai,t represents a vector of demoj
graphic characteristics which proxy for unobserved preferences a; ηi,t
are state-by-survey wave fixed
effects which proxy for unobserved prices, and ji,t captures remaining unobserved preferences.4 We
3 Given that this is a static framework, we abstract away from saving and so treat income y and expenditure x as
equal.
4 The canonical almost ideal demand system includes price indices and interactions between total consumption
5
estimate the set of all budget shares are simultaneously using a seemingly unrelated regression and
cluster standard errors at the village level. We use log total household income ln(yi,t ) as an instrument
for log total household expenditure to address possible measurement error.5
We also estimate a quadratic almost ideal demand system adapted from Banks, Blundell, and
Lewbell (1997):
j
wi,t
=
α1j
ln(xi,t ) +
α2j
2
ln(xi,t ) +
M
X
j
m
φjm zi,t
+ β j ai,t + ηi,t
+ ji,t
(4)
m=1
which conditions on quadratic log total expenditure and so allows a more flexible Engel curve. In
particular, this specification does not assume that preferences are homothetic and allows the budget
share to vary with income in a more general form. We again estimate the set of all budget shares
simultaneously using a seemingly unrelated regression, cluster standard errors at the village level, and
instrument the linear and quadratic expenditure terms with linear and quadratic income terms. As a
final robustness check, we estimate equations 3 and 4 with interactions between the log consumption
expenditure terms and the state-by-survey wave fixed effects. This allows the slope of the Engel curve
to vary with regional or temporal price variation, in the spirit of the full linear and quadratic almost
ideal demand systems.
The unitary model predicts that φjm = 0 for all j ∈ {1, . . . , J} and m ∈ {1, . . . , M }. The cooperative
model predicts that φjm /φjn = φkm /φkn for all j, k ∈ {1, . . . , J} and m, n ∈ {1, . . . , M }. We implement
these tests in section 4. We depart from the literature in estimating the tests for the entire sample and
separately for younger and older households. Our key innovation is the finding that the efficiency of
household resource allocations varies across households with different observed characteristics: younger
households’ allocations are not on average efficient while older households’ allocations are on average
efficient.
We also implement a second test based on Bourguignon, Browning, and Chiappori (2009). They
note that if there is at least one good dj and one observed distribution factor z m such that dj (x, z, p, a)
is monotone in z m , then dj can be inverted and z m written as a function of (x, z −m , p, a; dj ). This
function can be plugged into the budget share function for every other good k 6= j to yield a “zconditional demand function.” We therefore obtain
wk = θjk (x, z −m , p, a, ; dj )
expenditure and prices. Our data contain no direct measures of prices and prices can be calculated only for food shares.
We therefore omit the price terms except where noted below.
5 This strategy follows Attanasio, Battisin, and Mesnard (2011), who emphasize that total consumption expenditure
is typically calculated by aggregating over item-specific expenditure. Measurement errors in individual budget shares and
total consumption expenditure will thus be correlated by construction. Total income is likely to be a valid instrument
because it is reported separately on most surveys, so measurement error in this variable will be uncorrelated with
measurement error in the budget shares. The instrument passes standard weak instrument tests, with F-statistics of at
least 20 in all specifications.
6
from equation 2, where z −m = (z 1 , . . . , z m−1 , z m+1 , . . . z M ). Without loss of generality, we denote the
good and distribution factor used in the inversion by j = 1 and m = 1 respectively. We estimate this
equation using the linear specification
j
wi,t
= αj ln(xi,t ) +
M
X
j
m
1
φjm zi,t
+ β j ai,t + γ j wi,t
+ ηi,t
+ ji,t ,
(5)
m=2
j
j
where wi,t
, xi,t , ai,t , ηi,t
, and ji,t are defined as above. The φ vector contains the coefficients of interest.
Bourguignon, Browning, and Chiappori (2009) show that under the assumption of Pareto efficiency,
j
the distribution factors should have no direct effect on wi,t
for j > 2 after conditioning on w1 , the
expenditure share on good 1. Intuitively, w1 is a sufficient statistic for the effect of all the distribution
factors on all the budget shares. Hence, testing the cooperative model amounts to testing the joint
null hypothesis that φjm = 0 for all m ∈ {2, . . . , M } and for all j ∈ {2, . . . , J}
We also estimate quadratic versions of the z-conditional demand functions
j
wi,t
= α1j ln(xi,t ) + α2j ln(xi,t )2 +
M
X
j
m
1
φjm zi,t
+ β j ai,t + γ j wi,t
+ ηi,t
+ ji,t
(6)
m=2
which allow the slope of the Engel curve to vary nonlinearly over the observed support of income. We
again estimate equations 5 and 6 with a full set of interactions between the log consumption terms
and state-by-year fixed effects.
The z-conditional demand system permits a simple and direct test of the cooperative model, based
on a joint test of a vector of linear restrictions. In contrast, the joint proportionality test entails a
test of a nonlinear cross-equation set of restrictions. Bourguignon, Browning, and Chiappori (2009)
note that such tests have relatively low power and so may fail to reject the cooperative model even
when it does not apply. The z-conditional test does, however, have two key restrictions. First, the
test requires a monotonic relationship between at least one distribution factor and at least one budget
share. In our setting, this relationship needs to hold for both subsamples in which we implement
the cooperative test. Second, the test conditions on w1 which is by assumption correlated with the
unobserved preference term j for all j ∈ {2, . . . , J}. To address this problem, we use the excluded
distribution factor z 1 as an instrument for w1 in equations 5 and 6. This instrument is valid under the
assumptions used to derive the inversion argument above: w1 is a sufficient statistic for the effect of
all the distribution factors on all the budget shares. Hence, z 1 should affect wj for all j ∈ {2, . . . , J}
only through w1 . Instrument strength is testable, so we require that the relationship between z 1 and
w1 is both monotonic and statistically significant.
7
3
Distribution Factors, Expenditure and Demographics in PROGRESA 1998-1999
We use data from the evaluation of Mexico’s PROGRESA/Oportunidades program. This program,
started in 1997, still ongoing, and currently reaching about one quarter of the Mexican population,
provides conditional cash transfers to eligible households. To be eligible, a household must be sufficiently poor. The transfers, paid bimonthly, are largely in the form of scholarships to the last 4 grades
of primary school and the three grades of secondary school. These transfers are contingent on (i)
schoolchildren attending at least 85% of classes, (ii) household members undergoing periodical health
checks, and (iii) transfer recipients (typically, the mothers of the schoolchildren) attending nutrition
and health classes.
The evaluation data are collected from 506 rural villages across seven states whose school-aged
children met minimum school attendance and clinic visitation conditions. We have a complete census
of all village residents. Moreover, in our data, all households received a score on a wealth proxy-means
test: households who score below a threshold score are eligible for the PROGRESA transfers.
To evaluate the program, until the end of 1999 the transfers were offered only to eligible households
in 320 randomly chosen treatment villages. The baseline data was collected in September-October 1997,
while the first cash transfers were paid in March/April 1998 in the treated villages and between the
end of 1999 and the beginning of 2000 in the control villages. We also have three follow-up waves,
collected in October/November 1998, May/June 1999, and October/November 1999.6
Since the cash transfers were paid directly to mothers, women’s bargaining power in the household
plausibly increases among eligible households in treatment villages. Therefore, we follow the existing
literature (Bobonis, 2009; Angelucci and Attanasio, 2013; Attanasio and Lechene, 2014) and use an
indicator variable for Progresa eligibility as a distribution factor, denoted z 1 . Since the PROGRESA
cash transfer are conditional, they might change other features of the household which affect the
shape of the Engel curves, besides bargaining power. One relevant change that PROGRESA causes
is increased school attendance in households that would not have sent their children to school in the
absence of the program. This, in turn, may affect household demand by changing food and nonfood expenditures, as the children attending school may have school-related expenses and also change
their and their household members’ nutrition. Since primary school enrollment is virtually 100% in
these villages, the program can have minimal changes in primary school enrolment and attendance.
Conversely, secondary school enrollment in the absence of the program is about 65% and, indeed, the
program has been shown to have large effects on secondary school enrollment (Schultz, 2004; Angelucci,
de Giorgi, Rangel, and Rasul, 2010).
6 By November 1999, some eligible households in control villages had received their first transfer. We return to this
issue later.
8
To use living in a treated village as a distribution factor, we restrict the sample to households eligible
for PROGRESA in both treatment and control villages. Moreover, avoid potential confounding effects
of the program on households in which it caused a change in school attendance, we drop households
with potentially eligible secondary schoolchildren at baseline. To do that, we restrict the sample to
households without any child aged 10 to 16 with 5 or 6 maximum completed school grades in September
1997.7 These are the children who may start secondary school in 1998 or 1999, i.e. the group whose
enrollment is most likely affected by the program.89 Lastly, we drop single-headed households and also
all households for which we have missing observations. This provides a total of 25372 observations
spread across 506 villages and three survey waves.
Our second distribution factor is the municipality-level ratio of female to total population in 1995,
computed using data from the Mexican National System of Municipal Information database. This
variable is approximately continuous and a higher female:male sex ratio is likely to reduce the relative
bargaining power of women by skewing the marriage market in men’s favor. The distribution factors are assumed to be independent of prices and preferences. The second part of this independence
assumption may fail if there is selective migration out of villages in response to perceived marriage
market conditions. However, cross-state migration for marriage is not common in this setting: approximately 99% of husbands and 95% of wives live in their state of birth. Preliminary results show
that our main findings are robust to replacing the contemporaneous village sex ratio with the sex ratio
from the Mexican census closest to the time that each cohort reached age 18. This also goes some
way toward addressing the potential concern that the sex ratio at the time marriages are formed has
a larger impact on bargaining power within relationships than the contemporaneous sex ratio. The
latter measure is only appropriate if exit from marriages is a credible threat, as discussed in Mazzocco
(2007).
We split our sample into “old” and “young” households, defined by whether the household head is
above or below median age. Table 1 shows the means of socio-economic variables separately for the two
sub-samples (columns 1 and 2) and their difference (column 3). Older households have a bigger spousal
age gap and a higher likelihood of being indigenous, illiterate and less educated. They also have fewer
children aged 0-9, more children aged 10-18, and more adults in the household. Conversely, neither
the village characteristics – emigration share and marginalization index – nor the distribution factors
– the treatment dummy and female ratio – vary by sub-group. We also find that the characteristics of
these two groups are balanced by village type, with few exceptions, consistent with random assignment
(columns 4 and 5).
7 All results are robust to also dropping households with any children aged 10 to 16 irrespective of their level of
schooling at baseline.
8 Once enrolled in seventh grade, the first of the three years of secondary school, the likelihood of dropping out is
low.
9 The receipt of PROGRESA may also change health and nutrition for recipient households via changes in knowledge
and habits. However, Angelucci and Attanasio (2013) provide evidence inconsistent with this hypothesis using a sample
of urban program recipients.
9
Lastly, we regress female ratio on the socioeconomic variables by household sub-group (columns
6 and 7). As expected, female ratio is negatively correlated with spousal age gap (as one can marry
within one’s age range) and both positively correlated with village emigration (which occurs primarily
among males) and negatively correlated with village marginalization (as marginalized villages are
more geographically isolated, hence less likely to have emigrants). The other variables are generally
not statistically significantly correlated with female ratio, besides adult males and female, as expected.
Next, we repeat this exercise for household expenditure, income, and budget shares. We use
thirteen expenditure categories to construct the budget shares wj : seven food types (dairy, fats, fruit,
meat, starch, sugar, and vegetables) and six non-food commodities (children’s clothing, children’s
school-related expenditures, female clothing and household goods (eg. kitchen equipment), health,
male clothing and tobacco products, and and utilities).10
Table 2 shows that households are poor, as shown by their low income (about 1000 pesos at constant
1998 values, which equals approximately USD100), which is entirely consumed, and their budget is
spent largely on food. The main differences in budget shares between old and young households are
that the latter group, having younger children, spends more on child-related items. However, these
differences are not large. The other budget shares are similar. Even when there are statistically
significant differences, they are not very large.
When we look at the differences between treatment and control households, we find, as expected,
that expenditure increases in the treatment group, more so for younger households, which have more
school-age children and, therefore, larger cash transfers. Nevertheless, the effect of being in the treatment group on the budget shares is qualitatively similar for young and old households: the budget
shares of starchy food decreases while the budget shares of fruit and meat increase, as well as childrelated and women-related expenses, consistent with the evidence that women have more marked
preferences than men for private goods such as women’s clothing and household goods, and childrelated goods. The budget share for school and health expenses decrease. This change may be partly
caused by PROGRESA providing school- and health-related goods to its recipients.
The correlations between the female ratio and income, expenditures, and budget shares are also
qualitatively similar for old and young households. Areas with larger budget shares are wealthier
(higher income, consumption, and nonfood, meat, dairy, and utilities budgets shares; lower starches
budget share) and have a higher budget share on women, children, and health goods.
In sum, the descriptive relationship between consumption behavior and both distribution factors
is broadly consistent with the informal predictions of economic theory.
10 We consider several alternative expenditure categories as robustness checks. In addition to the seven food shares
used in the primary analysis, we consider four food shares based on the survey design (carbohydrates, fruit/vegetables,
protein, and other) and five food types following Attanasio and Lechene (2014) carbohydrates, fruit/vegetables, protein,
pulses, and other). In addition to the six non-food shares used in the primary analysis, we consider five shares with
children’s clothing and school-related expenditure combined together and four shares with all child-related expenditure
combined with female clothing and household goods. Results are robust to all nine combinations of these food and
non-food shares.
10
4
Tests of Efficiency in the Cooperative Model
4.1
Estimating Equations and Sample Definition
Recall that we estimate both conventional Engel curves of the form
j
wi,t
=
P
X
αpj ln(xi,t )p + α2j ln(xi,t )2 +
p=1
2
X
j
m
φjm zi,t
+ β j ai,t + ηi,t
+ ji,t
m=1
and z-conditional demand curves of the form
j
=
wi,t
P
X
j
2
1
αpj ln(xi,t )p + φj2 zi,t
+ ji,t .
+ β j ai,t + γ j wi,t
+ ηi,t
p=1
In each case we instrument log total expenditure ln(x) with log total income ln(y), estimate the budget
shares j = 1, . . . , J jointly as a system, and cluster the standard errors at the village level. We estimate
all models with four specifications: linear (P=1) and quadratic (P=2) terms in log expenditure and
with and without interactions between the log expenditure terms and state-by-survey wave fixed effects.
We use binary PROGRESA eligibility and continuous municipality-level sex ratio as distribution
factors. We estimate 13 budget shares: dairy, fats, fruit, meat, starch, sugar, vegetables, children’s
clothing, children’s school-related expenditures, female clothing and household goods (eg. kitchen
equipment), health, male clothing and tobacco products, and utilities. We include 18 covariates in
the a vector to proxy for preferences that affect the budget shares: household head age, education,
literacy, and indigenous status; counters for number of household members of each gender in the age
brackets 0-5, 6-9, 10-12, 13-15, 16-18, and ≥ 19; the village-level marginality index; and the share of
the village population who have migrated.
In the z-conditional demand estimation, we use the municipality sex ratio as the inverted distribution factor z 1 and utilities as the conditioning good w1 . The effect of the sex ratio on the budget share
spent on utilities is more robust than for other expenditure categories. We instrument the budget
share spent on utilities with the sex ratio. This exercise requires that (1) the sex ratio is a significant
determinant of the budget share spent on utilities and that (2) the relationship is monotonic. Both
assumptions are testable. The first stage F-statistic for the entire sample is at least 15.9 (p < 0.001)
across all specifications and the F-statistics for the samples of younger and older households are at
least 10.4 and 11.9 respectively (p ≤ 0.001). We test for monotonicity by including a the sex ratio
squared in the first stage and find that it is not statistically significant. We also note that the slope of
the utilities-sex ratio relationship implied by the quadratic estimates is monotone over the observed
support of the data. We cannot test whether the instrument is valid but this is implied by the maintained assumption of the model, that equilibrium w1 is a sufficient statistic for the effect of the sex
ratio on all budget shares.
11
We report results here including only households where no child is age-eligible to attend secondary
school. PROGRESA transfers depend in part on school attendance behavior, so the realized transfer will be endogeneous with respect to school attendance decisions. This concern is unlikely to be
important at primary school level, where enrollment is near universal. Schultz (2004) and Angelucci,
de Giorgi, Rangel, and Rasul (2010), amongst others, show that secondary school enrollment does respond to PROGRESA eligibility. All results are broadly robust to including households where one or
more members is age-eligible to attend secondary school in 1998/9 but has not yet completed primary
school and so is not grade-eligible.
4.2
Joint Proportionality Tests
Table 3 shows relevant results from estimating equation 3 with the full sample of households. PROGRESA eligiblity increases expenditure shares for fruit, meat and healthcare, while decreasing expenditure shares for education, utilities, children’s clothing and adult men’s clothing. This shift in expenditure may reflect the health and nutrition counselling required for PROGRESA recipients. Higher
values of the female:total population ratio are associated with increases in the expenditure shares for
vegetable and utilities, and with decreases in the expenditure shares for sugary foods and for adult
men’s and women’s clothing. Recall that higher values of this ratio are plausibly linked to lower female
bargaining power, so this pattern of results is perhaps surprising. Most non-food categories are luxury goods, whose expenditure share increases with income. This is consistent with a broad literature
documenting that low-income households spend a higher share of their income on food. Protein-rich
dairy and meat also appear to be luxury goods, while the expenditure shares of fats and starch are
falling in log total consumption.
The distribution factors are jointly significant in 8 of the 13 Engel curves. It is thus unsurprising
that the unitary model, which assumes that neither distribution factor have an effect on any budget
share, is strongly rejected. The χ2 test statistic for this set of exclusion restrictions is reported in
column 1 of panel A of table 5. The test statistic for the joint proportionality conditions implied by
the cooperative model is 20.07, which is significant at the 5% level. We therefore reject the hypothesis
that consumption allocations within households are on average Pareto efficient. Both results are highly
robust to a range of alternative model specifications. The test statistics are marginally larger when
controlling for a quadratic term in log total expenditure (columns 2 and 4) or interacting the log total
expenditure terms with state-by-survey wave fixed effects. This provides some reassurance that the
results are not driven by unobserved prices that prevent us from estimating a fully specified almost
ideal demand system.
Table 3 shows relevant results from estimating equation 3 separately for households whose head
is below the median age (panel A) or above the median age (panel B). The relationship between
budget shares and total expenditure is broadly consistent across the two subgroups: non-food shares
12
are generally luxuries, as are dairy and meat. Expenditure shares on fats and starch fall particularly
quickly as total expenditure rises.
The PROGRESA-eligible households in both age subsamples spend more on fruit, meat and healthcare, while decreasing expenditure on education and some categories of clothing. In both subsamples,
a higher fraction of women in the population is associated with much higher expenditure on utilities
and lower expenditure on sugary foods and adult men’s clothing. The magnitudes of many of these
relationships differ across the subsamples but the signs seldom differ and in very few cases can we
reject statistical equality across the two subsamples.
However, results of the formal tests reports in table 5 panels B and C do differ across the subsamples.
We reject the unitary model for both subsamples but by a considerably larger margin for the younger
than older households (χ2 test statistics 146 and 91 respectively, both with 12 degrees of freedom).
We reject the cooperative model for the younger households only, not for the older households. These
results are again robust to alternative model specifications, shown in columns 2-4. Some rejections
for the younger sample are marginal, in the sense that the p-values are in the neighborhood of 0.1.
We therefore turn to the z-conditional approach as an alternative, potentially more powerful, testing
strategy.
4.3
Z-Conditional Demand Tests
We find consistent test results using z-conditional demand systems, which are reported in table 6. We
reject the cooperative model for the entire sample across all specifications. The p-values for this test
are consistently smaller in the z-conditional test than the joint proportionaly tests discussed above,
reflecting the generally low power of cross-equation nonlinear tests. We reject the cooperative model
for younger households but fail to reject the cooperative model for older households. This result
is robust across all four specifications and, in results not reported here, to alternative definitions of
the expenditure categories and to alternative conditioning categories. The higher power of the zconditional tests is again visible in each subsample: the p-values for the cooperative test decrease
in the older sample from 0.49 to 0.19 and in the younger sample from approximately 0.10 to 0.04
(averaging across all four reported model specifications).
The z-conditional deman systems also provide a direct test of our hypothesis that the younger
and older subsamples differ in their consumption allocations. We test the statistical hypothesis that
the coefficients on the included distribution factor, PROGRESA eligibility, are equal for older and
younger households across all budget shares. This amounts to testing the economic hypothesis that
two samples’ consumption allocations across different goods responds in the same way to the resources
that PROGRESA eligibility potentially brings. This hypothesis is rejected across most specifications,
as shown in panel D of table 6. This direct test complements the indirect evidence of differences in
13
behavior obtained by comparing test results for younger and older households (inefficient vs efficient).11
We interpret these results as robust evidence that the cooperative model better describes the
intrahousehold bargaining process of households with older heads than households with younger heads.
In the next section we explore whether this difference is more consistent with cohort or life-cycle factors.
4.4
Reconciling Results with Attanasio and Lechene (2014)
Attanasio and Lechene (2014) (hereafter AL) test the collective model using the same dataset and fail
to reject the model. The difference between their and our results reflects several differences in our
implementation of the test. First, they use two waves of data after PROGRESA began to distribute
cash transfers; we use three waves of data. Second, they use the relative size of husbands’ and wives
family networks as a distribution factor; we use the local sex ratio. We prefer the local sex ratio
because it is “more continuous” than relative network sizes. Relative network sizes in our data range
from zero to one but 79.3% of the observed values are 0, 0.5, or 1. The local sex ratio ranges from
0.481 to 0.525 but no single value accounts for more than 7.2% of the data. Third, they model
only food consumption and condition on observed food prices as controls; we model both food and
non-food consumption and condition on state-by-survey wave fixed efects to proxy for unobserved
non-food prices. Fourth, we exclude households with high school-eligible children, whose enrollment
behavior determines households’ realized cash transfers. AL include these households and use distance
to the nearest secondary school as an instrument for enrollment. We have only recently obtained this
distance data and have not yet incorporated it into our main results. Fifth, they use village wages as
an instrument for household consumption; we use household income. In our sample the village wage
instrument does not pass standard instrument strength tests.
We replicate the AL result by estimating the same z-conditional demand system with the same two
waves of data, using the relative family network distribution factor, modeling only food consumption,
instrumenting high school enrollment with secondary school distance, and instrumenting household
expenditure with village wages. This yields substantively similar results: we fail to reject the collective
model with a Wald test statistic of 3.7 (p = 0.449). This result comes from inverting on the expenditure share on carbohydrates, which has a first-stage F -statistic of 13.5 (p < 0.001) with respect to
the relative family networks instrument. The first stages for secondary school enrollment and total
household expenditure have first-stage F -statistics of 10.0 and 44.3 with respect to school distance and
village wages respectively (p < 0.001 in both cases).
We also implement AL’s specifications using separate old and young subsamples. We find that
11
We are not aware of an analogous test available in the classic joint proportionality approach. This approach employs
j
joint nonlinear tests across the individual budget share equations of the form H0 :
χ2J−1
φ1
j
φ2
=
φk
1
∀j
φk
2
6= k. The resultant test
statistic has a
distribution under the null hypothesis. Taking the difference between the test statistics for younger
and older households does not yield useful insights, as there is no known closed-form expression for the difference between
two χ2 statistics.
14
the relative sibling network variable is not robustly significant in the budget share equations for these
subsamples. This makes the z-conditional testing framework infeasible when applying our subsample
comparison to AL’s specifications and datasets.
We interpret the difference between our and AL’s headline results as broadly consistent with our
observation that the validity of the collective model may be sample-specific. For example, including
households with high school-eligible children yields a marginal older sample (mean age 43.3 as opposed
to 42.7). AL’s sample thus includes more the older households whose consumption decisions we find
are more consistent with the collective model.
5
Is Inefficiency a Cohort or Life Cycle Phenomenon?
This difference in the efficiency of consumption decisions between older and younger households is
consistent with both cohort and life-cycle effects. Households may reach more efficient outcomes over
repeated rounds of bargaining, perhaps by learning about each others’ preferences or because exit
becomes a less credible threat for at least one partner. Alternatively, efficiency may be an essentially
time-varying characteristic of households that is more common amongst the cohort of rural Mexican
households with older heads. One obvious life-cycle explanations can be quickly tested: older individuals are more likely to be divorced or separated than younger individuals but this difference is relatively
small (2.8% versus 1.9%). It is possible that households that reach inefficient allocations are more
likely to separate over time but the relatively small number of separations in rural Mexico makes this
explanation unlikely.
We attempt to separate cohort and life-cycle effects in two ways. First, we consider characteristics
of the older and younger cohorts that may drive the difference in the efficiency of household allocations.
If there exist characteristics that (1) do not vary through the life-cycle but (2) differ across cohorts
and (3) are associated with cross-sectional differences in the efficiency of household allocations, this
provides evidence in support of cohort factors. Second, we implement the young-old efficiency tests
using data from a follow-up survey of the same villages in 2007. The age distribution of household heads
remains approximately constant between 1998/9 and 2007 but there is some change in the distribution
of birth cohorts. If the results are broadly consistent between 1998/9 and 2007, this provides evidence
in support of life-cycle factors.
5.1
Efficiency Comparisons Across Time-Invariant Cohort Characteristics
Here we consider characteristics of the older and younger cohorts might potentially drive the difference
in the efficiency of household allocations. If there exist characteristics that (1) do not vary through
the life-cycle but (2) differ across cohorts and (3) are associated with cross-sectional differences in
the efficiency of household allocations, this provides evidence in support of cohort ahead of life-cycle
15
factors. We depart from the finding that in households with high scores on a “tradition” index,
PROGRESA eligibility is more likely to cause rising domestic violence Angelucci (2008). This reflects
anthropological evidence that increases in female income share within the household are associated
with disruptions to men’s perceptions of their traditional gender roles. Angelucci (2008) constructs
an index of households’ tradition-orientation by averaging three proxies: the age gap between spouses,
the husband’s education (which enters the index negatively), and the husband’s age. We construct
this index, split households into those with above- and below-median values, and implement the same
analysis described above.
The test results are shown in tables 7 and 8. The main patterns are fairly similar to those shown
in tables 5 and 8. The unitary model is rejected for both subsamples but by a larger margin for the
less than more traditional sample. When using the joint proportionality test (table 7), the cooperative
model is rejected for the less traditional but not for the more traditional sample. However, the
cooperative model is rejected for both subsamples when using the z-conditional tests (table 8), albeit
by a significantly larger margin for the less traditional households.
The tradition and household head age splits are fairly strongly correlated (ρ = 0.69) but there
is substantial variation in the former measure conditional on household head age. We also construct
a tradition index using only the two time-invariant characteristics, the husband’s education and the
spousal age gap. The resultant split is less strongly correlated with the split on household head age
(ρ = 0.44). We find similar results using this entirely time-invariant tradition index. The joint proportionality tests reject the cooperative model for less traditional households but not more traditional
households (p = 0.004 and 0.609 respectively), while the z-conditional tests reject for both subsamples
but by significantly different margins (p = 0.000 and 0.035 respectively, with p = 0.049 for equality
across subsamples).
These results show that cohort-level characteristics that vary little over the life-cycle are also
associated with differences in the extent to which consumption allocations are consistent with the
cooperative model. However, the distinction is somewhat less clear for these time-invariant cohort
characteristics than for the household head age. We do not believe that this provides strong evidence
that the age pattern observed in the cross-section can be ascribed to either cohort or life-cycle factors.
We therefore turn to an alternative testing strategy that employs a later survey of the same villages.
5.2
Efficiency Comparisons between 1998/9 and 2007
The main results are based on three sequential censuses of households in sampled villages in 1998 and
1999, from which we construct a household panel. The same villages were surveyed again in 2003
and 2007, albeit using different survey instruments. In preliminary results, we test the validity of the
cooperative model using the 2007 wave and compare these results to the 1988/9 wave. The 2007 sample
16
is slightly older than the 1998/9 sample and was born slightly later.12 If older households attain more
efficient consumption allocations due to life-cycle factors, then allocations should be more efficient for
the sample in 2007, which is relatively older. If older households attain more efficient consumption
allocations due to cohort factors, then allocations should be less efficient for the sample in 2007, which
was born in later cohorts.
The main empirical strategy requires some modifications when applied to the 2007 data. In particular, there is experimental variation in village-level PROGRESA eligiblity in 1998/9 due to random
rollout of the program. By 2007, all surveyed villages are eligible for transfers via PROGRESA. We
therefore construct an alternative distribution factor based on the share of total household income
controlled by the wife: her labor income and transfers paid to her, including from PROGRESA. This
provides a proxy for the wife’s bargaining power in the household and ranges from 0 to 1 with mean
0.61 and interquartile range 0.26-1.00. PROGRESA transfers constitute by far the largest portion
wives’ income share. The cross-household variation in income share is explained by differences in
PROGRESA eligibility (grants depend on the number and gender composition of relevant children)
and take-up, as well as differences in labor force participation. This variation may be correlated with
unobserved household-level preferences and so wife’s income share in the 2007 data is on balance a
less compelling distribution factor than the randomly assigned PROGRESA eligibility in the 1998/9
data. We use municipality sex ratios in 2005 as the other distribution factor. We once again use
the sex ratio for inverting the z-conditional demand functions, expenditure share on utilities as the
conditioning good, and the wife’s income share to conduct the cooperative model test.
In addition, the survey only asks education levels for a subset of respondents so it is impossible to
obtain a comprehensive measure of husbands’ education. We therefore report test results for subsamples defined by husbands’ age and by the spousal age gap in tables 9 and 10 respectively. Finally, the
survey asks about school enrollment status for a subset of households. We therefore cannot identify
households with children who are enrolled or are eligible to enroll in secondary school and so cannot
remove them from the estimation sample.
The tables present three key patterns in the 2007 data. First, the cooperative model is strongly
rejected for the entire sample. The test statistics are comparable in magnitude to those estimated
using 1998/9 data, despite the substantially smaller sample size (18834 households in 2007 versus
25392 households in 1998/9).13 The cooperative model is also rejected for households above and below
the median head age and, in most specifications, above and below the median spousal age gap. The key
dimension of heterogeneity observed in the 1998/9 is substantially less prominent in the 2007, though
we continue to reject equality of effect of wife’s income share on budget shares across the younger and
12 The mean ages in 2007 and 1998/9 are respectively 46.7 and 43.8 and the interquartile ranges are [35,58] and [32,53].
The mean birth cohorts in 2007 and 1998/9 are respectively 1962 and 1957 and the interquartile ranges are [1949,1972]
and [1945,1967].
13 The test statistics are asymptotically χ2 -distributed so it is unclear whether their finite sample values should be
adjusted to reflect the differences in sample size.
17
older households.14
The 2007 results show stronger evidence against the cooperative model than the 1998/9 results
and less evidence of heterogeneity. We interpret this preliminary cross-period result as evidence that
intrahousehold consumption allocations may be less efficient amongst older households who are members of later cohorts than amongst younger households who are members of earlier cohorts. This shifts
our interpretation of the main results, which employ within-period comparisons between younger and
older households, toward a cohort effect rather than a life-cycle effect. Households headed by individuals born in later years appear less likely to achieve Pareto efficient intrahousehold consumption
allocations.
6
Public Good Provision: Children’s Education
Inefficient bargaining can have important welfare implications, including under-provision of public
goods Lundberg and Pollak (1993). Our analysis provides a unique opportunity to test this, as we
identify subsamples that contain different proportions of efficient households but face similar prices
and institutional environments. We proceed by testing whether children’s education attainment and
participation, both plausible measures of household public goods, are lower in the subsample of younger
households, for whom we reject collective rationality. We document that children in younger households
have lower levels of education attainment before receiving cash transfers and their education attainment
and participation are less response to receipt of cash transfers.
We estimate the parameters of the following equation for household i at time t:
yit = α0 + α1 tt + α2 Pi + α3 oi + α4 tt Pi + α5 tt oi + α6 Pi oi + α7 tt Pi oi + βXit + it ,
(7)
where the indicator variables t, P , and o equal 1 for respectively post-treatment periods, eligible
households in Progresa villages, and households with older than median husbands. This is effectively a
triple-difference model, comparing the time change in the outcome variables for eligible and ineligible
and older and younger households. We consider four outcome variables y: (1) the share of offspring of
the household head aged 11 to 16 with at least completed primary education, (2) the share of offspring
of the household head aged 11 and older with at least completed secondary education, (3) the share
of children aged 6-11 enrolled in school, and (4) the share of children aged 11-16 who are enrolled in
school. We choose the age brackets 6-11 and 11-16 because primary school lasts 6 years and children
often first enroll at age 5. Therefore, children as young as 11 can complete primary school. Moreover,
at baseline, a large fraction of children without complete primary school are not enrolled by age 16.
We include in X the share of children in the household aged 6 to 16 (one variable per year of
14 This result may not be robust to using joint proportionality tests, but those results were not yet available at the
time this draft was prepared.
18
age), the village marginalization index, and state fixed effects. The parameters of interest are α3
and α3 + α6 , which measure baseline differences in the educational attainment of children from old
and young households in control and treatment villages, and α7 , which measures whether the average
treatment effect of being an eligible household in a Progresa village differs for children from old and
young households.
We report the estimates of these parameters in Table 11.15 We report three sets of estimates for
each outcome. The first drops households with any child aged 10 to 16 with 5 or 6 maximum completed
school grades in September 1997. This is the sample we use in the rest of the analysis. The second
column keeps the same sample and adds the household wealth index and the education and indigenous
status of the household head, measured at baseline, to see whether any differences between old and
young households are explained by differences in these variables, as older households are less educated
and more likely to be indigenous. In this way, we control for potential differences between young and
old households caused by different preferences or credit constraints. Lastly, the third column shows the
results for the sample without dropping households with any child aged 10 to 16 with 5 or 6 maximum
completed school grades in September 1997.
Rows 1 and 2 in columns 1 to 6 show that, at baseline, the share of children with at least completed
primary (secondary) school is between 1.6 and 4.9 (1.3 and 3.5) percentage points higher in old than in
young households in control villages and that this difference is not explained by differences in wealth,
education, and indigenous status (columns 2 and 5). These differences are economically large, as the
baseline means in the control group are 0.56-0.51 for primary school and 0.07-0.16 for secondary school
in the restricted and non restricted samples.
Rows 3 and 4 in columns 1 to 6 further shows that Progresa eligibility has no effect on the share
of children with at least completed primary school in younger households, but there is some evidence
that this effect is in fact statistically significant and 3 percentage point higher. Conversely, Progresa
eligibility has no effect for the share with completed secondary school for both young and old households. This is consistent with the evidence that the strongest effects of Progresa on education are on
continuing to attend school beyond primary education: if the program induces children who would
have stopped at sixth grade to enroll in 7th grade, then we would not expect treatment effects on
complete secondary schooling until 2000, that is at least 3 years since the program start.
Rows 1 and 2 in columns 7 to 12 show that, at baseline, the shares of children aged 6-11 (12-16)
enrolled in school are up to 2.9 (6.4) percentage points lower in old than in young households. This
is consistent with the higher stock of education: as children are more likely to complete primary and
secondary education in old households, they are, therefore, less likely to be enrolled in school.
15 All results are robust to conditioning on measures of baseline socio-economic status: the household wealth index,
the head of household education, knowledge of indigenous languages, spousal age gap, an indicator for couple-headed
households. Therefore, the differences in educational attainment at baseline and in response to the eligibility to Progresa
are not explained by these characteristics.
19
Rows 3 and 4 in columns 7 to 12 further shows that Progresa eligibility has a statistically and
economically larger effect on school enrollment, increasing up to 3.8 percentage points more in old
than in young households, although this difference is statistically significant only for enrollment for
the younger children.
These results provide evidence of under-investment in public goods, specifically of children’s human capital, for the young subsample that contains fewer efficient households. These findings have
important implications for the optimal design of public policy. Differential household responses to the
same policy can have important welfare effects that might require more tailored policies. Conditional
cash transfers have become a very popular policy tool in many countries but their effectiveness will be
reduced if inefficient household bargaining leads to under-investment in education. At the margin, this
might increase the attractiveness of policies that provide transfers to households that cannot be lost
due to inefficient household bargaining. For example, school tuition fee reductions can produce similar
enrollment effects to conditional-on-enrollment cash transfers, without providing a direct pecuniary
transfer to households that might be inefficiently spent (Garlick, 2013). Note that this issue applies
to the Pareto efficiency of how cash transfers are spent, not to the adherence of household spending
with policymakers’ preferences.
These findings have important policy implications because they indicate strong differential responses to the same policy, correlated with differences in household bargaining.
7
Conclusion
We present a novel empirical finding regarding the validity of the cooperative model of household
resource allocation. This model predicts that households containing members with potentially heterogeneous preferences will engage in a process of bargaining that yields a Pareto efficient set of resource
allocations. Prior empirical evidence regarding this model is mixed: most but not all tests fail to
reject the key testable prediction. We show that within a sample of rural Mexican households, Pareto
efficiency of household expenditure is rejected for younger but not older households. This result is robust to a range of model specifications, sample definitions, and to using more aggregated expenditure
categories, though the latter results are not presented here.
Our finding suggests that tests for “the validity of cooperative household models” may be misplaced;
researchers should instead aim to identify which models best characterize different groups of households.
For example, the disparate results of prior empirical tests may reflect differences in the composition
of samples. Standard approaches that test whether an entire sample’s behaviour is consistent with
a specific model do not take this into account. We present preliminary evidence that this finding
may be driven by a time-invariant cohort-level differences between younger and older houseolds. This
will ultimately help to identify the factors that explain why some households achieve Pareto-efficient
20
resource allocations and others do not. This contributes to an active research program into the best
way to model non-cooperative household behavior.
References
Akresh, R. (2005): “Understanding Pareto Inefficient Intrahousehold Allocations,” Discussion Paper
1858, IZA.
Angelucci, M. (2008): “Love on the Rocks: Domestic Violence and Alcohol Abuse in Rural Mexico,”
The B.E. Journal of Economic Analysis and Policy: Contributions, 8(1).
Angelucci, M., and O. Attanasio (2013): “The Demand for Food of Poor Urban Mexican Households: Understanding Policy Impacts using Structural Models,” American Economic Journal: Policy, 5(1), 146–178.
Angelucci, M., G. de Giorgi, M. Rangel, and I. Rasul (2010): “Family Networks and School
Enrollment: Evidence from a Randomized Social Experiment,” Journal of Public Economics, 94(34), 197–221.
Ashraf, N. (2009): “Spousal Control and Intra-Household Decision Making: An Experimental Study
in the Philippines,” American Economic Review, 99(4), 1245–1277.
Attanasio, O., E. Battisin, and A. Mesnard (2011): “Food and Cash Transfers: Evidence from
Colombia,” Economic Journal, 122(559), 92–124.
Attanasio, O., and V. Lechene (2014): “Efficient Responses to Targeted Cash Transfers,” Journal
of Political Economy, 122(1), 178–222.
Banks, J., R. Blundell, and A. Lewbell (1997): “Quadratic Engel Curves and Consumer Demand,” Review of Economics and Statistics, 79(4), 527–539.
Bobonis, G. (2009): “Is the Allocation of Resources within the Household Efficient? New Evidence
from a Randomized Experiment,” Journal of Political Economy, 117(3), 453–503.
Bourguignon, F., M. Browning, and P.-A. Chiappori (2009): “Efficient Intra-Household Allocations and Distribution Factors: Implications and Identification,” Review of Economic Studies,
76(2), 503–528.
Bourguignon, F., M. Browning, P.-A. Chiappori, and V. Lechene (1993): “Intra Household
Allocation of Consumption: A Model and Some Evidence from French Data,” Annales d’Economie
et de Statistique, 29, 137–156.
Browning, M., F. Bourguignon, P.-A. Chiappori, and V. Lechene (1994): “Income and
Outcomes: A Structural Model of Intrahousehold Allocation,” Journal of Political Economy, 102(6),
1067–1092.
Browning, M., and P.-A. Chiappori (1998): “Efficient Intra-household Allocation: A General
Characterization and Empirical Tests,” Econometrica, 66(6), 1241–1278.
Cherchye, L., B. D. Rock, and F. Vermeulen (2009): “Opening the Black Box of Intrahousehold
Decision Making: Theory and Nonparametric Empirical Tests of General Collective Consumption
Models,” Journal of Political Economy, 117(6), 1074–1104.
Chiappori, P.-A. (1992): “Collective Labor Supply and Welfare,” Journal of Political Economy,
100(3), 437–467.
21
Chiappori, P.-A., B. Fortin, and G. Lacroix (2002): “Marriage Market, Divorce Legislation and
Household Labor Supply,” Journal of Political Economy, 110(1), 37–72.
Deaton, A., and J. Muellbauer (1980): “An Almost Ideal Demand System,” American Economic
Review, 70(3), 213–326.
Dercon, S., and P. Krishnan (2000): “In Sickness and in Health: Risk Sharing within Households
in Rural Ethiopia,” Journal of Political Economy, 108(4), 688–727.
Djebbari, H. (2005): “The Impact on Nutrition of the Intrahousehold Distribution of Power,” Discussion Paper 1701, IZA.
Duflo, E., and C. Udry (2004): “Intrahousehold Resource Allocation in Cote d’Ivoire: Social
Norms, Separate Accounts and Consumption Choices,” Working Paper 10498, NBER.
Farfán, G. (2013): “Extended Families across Mexico and the United States,” Duke University.
Garlick, R. (2013): “How Price Sensitive is Primary and Secondary School Enrollment? Evidence
from Nationwide Tuition Fee Reforms in South Africa,” Mimeo, University of Michigan.
Hoel, J. (2013): “When the Average Obscures What is True for Most: A Within-Subject Test of
Household Model Assumptions,” International Food Policy Research Institute.
Lundberg, S., and R. Pollak (1993): “Separate Spheres Bargaining and the Marriage Market,”
Journal of Political Economy, 101(6), 988–1010.
Mazzocco, M. (2007): “Household Intertemporal Behavior: A Collective Characterization and a
Test of Commitment,” Review of Economic Studies, 74(3), 857–895.
Rangel, M. (2006): “Alimony Rights and Intrahousehold Allocation of Resources: Evidence from
Brazil,” Economic Journal, 116(513), 627–658.
Rangel, M., and D. Thomas (2005): “Out of West Africa: Evidence on the Efficient Allocation of
Resources within Farm Households,” University of Chicago.
Schaner, S. (2013): “Do Opposites Detract? Intrahousehold Preference Heterogeneity and Inefficient
Strategic Savings,” Dartmouth College.
Schultz, P. (2004): “School Subsidies for the Poor: Evaluating the Mexican PROGRESA Poverty
Program,” Journal of Development Economics, 74(1), 199–250.
Thomas, D. (1990): “Intra-household Resource Allocation: An Inferential Approach,” Journal of
Human Resources, 25(4), 635–664.
Udry, C. (1996): “Gender, Agricultural Production, and the Theory of the Household,” Journal of
Political Economy, 104(5), 1010–1046.
22
Table 1: Summary Statistics and Balance Tests for Households by Age Group: Socio-Economic Status
Old
Young
O - Y
Variable
T-C | O
T-C | Y
Female ratio
Old
Young
Coeff.
[st.err.]
-42.048
-16.247
[15.100]***
[8.360]*
8.819
-8.327
[23.967]
[9.997]
-5.429
-12.104
[6.672]
[8.088]
-3.866
-4.643
[2.890]
[2.750]*
0.555
-0.096
[1.633]
[1.112]
-21.924
0.711
[17.411]
[12.533]
-1.172
-2.178
[1.085]
[1.358]
0.077
-0.729
[1.256]
[1.352]
-2.477
-0.207
[0.945]***
[0.953]
-0.632
-0.335
[0.935]
[1.044]
-1.116
0.825
[0.873]
[0.573]
-0.782
-0.68
[0.718]
[0.591]
-0.788
0.686
[0.784]
[0.391]*
-0.928
-0.217
[0.670]
[0.403]
-1.169
-0.243
[0.734]
[0.369]
0.524
-0.168
[0.744]
[0.422]
-3.19
-0.733
[1.780]*
[0.452]
2.465
1.918
[1.587]
[0.632]***
0.581
1.118
[0.581]***
[0.430]**
-11.376
-13.360
[4.987]**
[4.876]***
2.500
-4.261
[3.417]
[3.111]
1
1
Mean
Coeff.
Coeff.
(st. dev.)
[st.err.]
[st.err.]
spousal age gap
6.192
2.838
3.354
-0.283
-0.092
(7.651)
(4.382)
[0.154]***
[0.309]
[0.159]
husband age
58.084
29.256
28.828
-0.148
-0.033
(12.368)
(5.166)
[0.246]***
[0.467]
[0.209]
husband education
1.665
4.102
2.438
0.187
-0.192
(1.923)
(2.381)
[0.074]***
[0.122]
[0.162]
husband indigenous
0.432
0.356
0.076
0.007
0.009
(0.495)
(0.479)
[0.017]***
[0.058]
[0.057]
husband literate
0.569
0.836
-0.267
-0.067
0.018
(0.495)
(0.37)
[0.013]***
[0.030]**
[0.022]
# relatives
2.771
2.777
-0.006
-1.237
0.121
(4.837)
(3.162)
[0.212]
[0.505]**
[0.204]
#males aged ≤ 5
0.31
0.765
-0.456
-0.018
0.04
(0.619)
(0.782)
[0.017]***
[0.023]
[0.028]
# females aged ≤ 5
0.321
0.783
-0.463
0.025
-0.044
(0.644)
(0.791)
[0.017]***
[0.025]
[0.027]
# males aged 6 - 9
0.263
0.395
-0.132
0.003
0.038
(0.53)
(0.617)
[0.012]***
[0.019]
[0.019]**
# females aged 6 - 9
0.249
0.369
-0.120
0.009
-0.014
(0.515)
(0.608)
[0.012]***
[0.018]
[0.020]
# males aged 10-12
0.184
0.129
0.055
-0.009
-0.001
(0.433)
(0.366)
[0.009]***
[0.015]
[0.013]
# females aged 10-12
0.169
0.123
0.046
0.004
-0.017
(0.414)
(0.356)
[0.008]***
[0.014]
[0.012]
# males aged 13-15
0.188
0.064
0.123
0.02
-0.003
(0.424)
(0.27)
[0.008]***
[0.014]
[0.009]
# females aged 13-15
0.165
0.061
0.104
-0.034
-0.017
(0.408)
(0.257)
[0.007]***
[0.015]**
[0.009]**
# males aged 16-18
0.18
0.042
0.138
0.027
-0.015
(0.42)
(0.213)
[0.008]***
[0.014]*
[0.007]**
# females aged 16-18
0.174
0.072
0.102
0.002
0.006
(0.418)
(0.267)
[0.007]***
[0.013]
[0.008]
# males aged ≥ 19
1.474
1.041
0.433
-0.002
0.015
(0.73)
(0.266)
[0.015]***
[0.032]
[0.009]*
# females aged ≥ 19
1.466
1.047
0.419
-0.001
0.009
(0.737)
(0.423)
[0.015]***
[0.030]
[0.018]
Village emigration share
0.048
0.043
0.005
0.008
0.001
(0.073)
(0.063)
[0.006]
[0.009]
[0.008]
Village marginalization
0.465
0.464
0.001
-0.020
0.011
index
(0.685)
(0.781)
[0.066]
[0.094]
[0.101]
Treatment
0.661
0.607
0.055
1
1
(0.474)
(0.490)
[0.043]
Female ratio 1995
0.501
0.501
0.001
0.001
-0.002
(0.009)
(0.010)
[0.001]
[0.001]
[0.001]
Notes: The sample consists of 25372 household-by-survey wave observations in 506 villages.
23
Table 2: Summary Statistics and Balance Tests for Households by Age Group: Socio-Economic Status
and Demographics
Old
Young
T-C | O
O - Y
T-C | Y
Female ratio
Old
Young
Variable
Mean
Coeff.
Coeff.
Coeff.
(st. dev.)
[st.err.]
[st.err.]
[st.err.]
income
1037.016
948.599
88.417
44.706
-121.403
3,629.25
4,732.06
(2829.693)
(5927.015)
[60.741]
[69.806]
[149.072]
[3,364.515]
[2,991.726]
expenditure
1002.558
997.293
5.265
49.869
116.335
4,709.38
4,475.08
(932.802)
(1069.168)
[18.539]
[29.552]*
[34.328]*** [1,757.944]***
[1,937.912]
food share
82.688
81.529
1.159
0.442
0.376
-151.464
-217.041
(12.717)
(11.944)
[0.233]***
[0.533]
[0.636]
[28.056]***
[30.206]***
nonfood share
17.312
18.471
-1.159
-0.442
-0.376
151.464
217.041
(12.717)
(11.944)
[0.233]***
[0.533]
[0.636]
[28.056]***
[30.206]***
starches share
42.292
41.636
0.656
-0.097
-1.401
-118.085
-156.753
(14.948)
(14.48)
[0.276]**
[0.637]
[0.643]**
[30.250]***
[32.117]***
vegetable share
7.72
7.839
-0.119
-0.036
0.156
3.315
9.359
(5.519)
(5.537)
[0.092]
[0.181]
[0.172]
[10.129]
[8.543]
fruit share
2.688
2.764
-0.076
0.384
0.583
5.679
-8.247
(4.271)
(4.333)
[0.067]
[0.157]**
[0.160]***
[8.532]
[9.283]
meat share
7.353
6.96
0.393
0.509
1.24
31.418
22.41
(9.084)
(8.425)
[0.173]**
[0.396]
[0.354]***
[17.502]*
[16.935]
dairy share
6.137
6.76
-0.623
0.055
0.209
29.874
30.941
(6.281)
(6.581)
[0.103]***
[0.215]
[0.247]
[10.824]***
[13.344]**
sugar share
10.757
9.839
0.919
-0.259
-0.117
-94.526
-101.814
(8.407)
(7.764)
[0.140]***
[0.354]
[0.344]
[18.027]***
[16.254]***
fats share
5.74
5.732
0.009
-0.114
-0.295
-9.139
-12.938
(4.846)
(5.223)
[0.080]
[0.151]
[0.150]*
[7.512]
[7.509]*
women share
1.17
1.185
-0.015
0.135
0.145
14.461
9.206
(2.241)
(2.115)
[0.034]
[0.072]*
[0.068]**
[3.604]***
[3.185]***
men share
1.244
1.097
0.146
0.104
0.054
5.133
7.316
(2.763)
(2.247)
[0.036]***
[0.083]
[0.070]
[4.898]
[3.290]**
kids share
1.408
2.644
-1.236
0.338
0.604
4.551
28.919
(2.566)
(3.283)
[0.061]***
[0.084]***
[0.130]***
[4.394]
[5.981]***
school share
1.242
1.364
-0.123
-0.23
-0.26
-3.062
15.757
(4.144)
(3.86)
[0.064]*
[0.116]**
[0.127]**
[5.907]
[5.867]***
health share
8.109
7.953
0.158
-0.422
-0.579
66.88
52.244
(9.167)
(7.598)
[0.122]
[0.257]
[0.249]**
[12.747]***
[13.126]***
utilities share
4.139
4.228
-0.089
-0.367
-0.34
63.5
103.6
(5.094)
(5.042)
[0.104]
[0.254]
[0.283]
[13.112]***
[13.781]***
Notes: The sample consists of 25372 household-by-survey wave observations in 506 villages.
24
25
Fats
-3.025***
(0.578)
0.043
(0.110)
8.210
(6.199)
1.852
Fruit
0.023
(0.500)
0.476***
(0.122)
0.341
(8.268)
15.611***
Meat
5.076***
(1.068)
0.590**
(0.268)
-15.743
(14.972)
6.065**
Starch
-13.069***
(1.929)
0.204
(0.479)
-40.614
(31.703)
2.002
Sugar
-0.330
(0.953)
0.117
(0.250)
-28.584**
(12.587)
6.291**
Vegetables
0.826
(0.682)
-0.036
(0.140)
14.256*
(7.540)
3.757
Children
-2.985***
(0.930)
-0.397**
(0.175)
13.798
(9.553)
7.740**
Table 3: Engel Curves for Full Sample
Health
2.234***
(0.320)
0.260***
(0.061)
-0.818
(4.151)
18.166***
Men
2.262***
(0.272)
-0.122**
(0.056)
-7.451**
(2.943)
8.989**
School
0.866*
(0.455)
-0.292***
(0.101)
4.957
(5.505)
9.747***
Utilities
2.632***
(0.643)
-0.651***
(0.170)
41.806***
(9.822)
28.160***
Women
1.687***
(0.260)
-0.015
(0.050)
-4.015*
(2.474)
2.756
χ2 test stat. for
both dist. factors
Mean of dep. var.
6.453
5.736
2.727
7.153
41.959
10.291
7.780
8.030
2.036
1.169
1.304
4.184
4.517
Notes: The sample consists of 25372 household-by-survey wave observations in 506 villages. Standard errors in parentheses are clustered at the village level. *, **, and *** denote
significance at the 10, 5, and 1% levels respectively. The system of equations is estimated simultaneously in a seeming unrelated regression model. Log household expenditure instrumented
by log household income with first stage F-statistic 225.95 (p < 0.001). All equations include state-by-survey wave fixed effects; controls for household head age, education, literacy,
and indigenous descent; controls for the number of household members of each gender in ages brackets 0-5, 6-9, 10-12, 13-15, 16-18, and 19 or older; and controls for the village-level
marginalization index and migrant share.
Log HH
expenditure
PROGRESA
eligibility
Sex ratio
Dairy
3.804***
(0.719)
-0.176
(0.153)
13.946
(9.095)
3.702
26
Table 4: Engel Curves for Sample Split by Age of Household Head
Panel A: Households with Younger Heads
Dairy
Fats
Fruit
Meat
Starch
Sugar
Vegetables
Children
Health
Men
School
Utilities
Women
Log HH
5.759***
-2.750**
-0.754
6.289***
-19.410***
-2.194
0.461
-1.804
2.802***
3.430***
1.286*
4.702***
6.271***
expenditure
(1.335)
(1.155)
(0.868)
(1.985)
(3.453)
(1.577)
(1.186)
(1.596)
(0.657)
(0.496)
(0.726)
(1.053)
(1.308)
PROGRESA
-0.384
0.031
0.578***
0.488
0.885
0.408
0.068
-0.481*
0.209**
-0.328***
-0.392***
-0.958***
-0.305
eligibility
(0.107)
(0.164)
(0.150)
(0.343)
(0.581)
(0.306)
(0.194)
(0.250)
(0.101)
(0.068)
(0.135)
(0.209)
(0.207)
Sex ratio
8.807
8.018
-5.388
-21.649
-42.260
-25.310*
13.906
6.846
2.587
-7.197**
14.818**
53.847***
10.362
(12.708)
(7.349)
(9.523)
(17.200)
(34.003)
(14.319)
(8.842)
(11.368)
(5.826)
(2.961)
(6.243)
(11.356)
(10.904)
χ2 test stat. for
3.099
1.192
14.787***
3.780
4.261
6.431**
2.503
4.477
4.321
25.256***
20.800***
41.295***
3.724
both dist. factors
Mean of dep. var.
6.760
5.732
2.764
6.960
41.654
9.839
7.839
7.953
2.644
1.097
1.364
4.228
5.193
Panel B: Households with Older Heads
Dairy
Fats
Fruit
Meat
Starch
Sugar
Vegetables
Children
Health
Men
School
Utilities
Women
Log HH
2.964***
-2.982***
0.151
4.231***
-8.584***
0.440
1.097
-4.135***
1.590***
1.553***
0.605
1.659**
3.606***
expenditure
(0.802)
(0.607)
(0.592)
(1.147)
(2.120)
(1.133)
(0.770)
(1.127)
(0.288)
(0.326)
(0.584)
(0.679)
(0.788)
PROGRESA
-0.145
0.018
0.427
0.589***
0.013*
0.006
-0.051
-0.396*
0.264***
-0.008
-0.230*
-0.534***
0.093
eligibility
(0.158)
(0.140)
(0.130)
(0.327)
(0.545)
(0.284)
(0.160)
(0.218)
(0.057)
(0.072)
(0.120)
(0.180)
(0.567)
Sex ratio
19.808**
8.598
9.216
-12.028
-34.052
-30.622*
13.878
21.272*
-5.293
-9.064**
-4.995
25.030**
-12.035
(9.242)
(8.167)
(8.874)
(17.214)
(33.957)
(15.666)
(9.221)
(11.899)
(3.501)
(4.180)
(6.197)
(9.939)
(9.163)
2
χ test stat. for
6.446**
1.157
12.704***
3.700
1.065
3.845
2.354
6.867**
25.995***
4.847*
4.895*
12.738***
1.988
both dist. factors
Mean of dep. var.
6.137
5.740
2.688
7.353
42.292
10.757
7.720
8.110
1.408
1.244
1.242
4.139
3.819
Notes: Sample consists of 25372 household-by-survey wave observations in 506 villages. The young and old subsamples consist of 12888 and 12484 household-by-survey wave osbervations
respectively. Standard errors in parentheses are clustered at the village level. *, **, and *** denote significance at the 10, 5, and 1% levels respectively. The system of equations,
for young and old households, is estimated simultaneously in a seeming unrelated regression model. Log household expenditure instrumented by log household income with first stage
F-statistics 80.32 (p < 0.001) and 185.27 (p < 0.001) for the young and old subsamples respectively. All equations include state-by-survey wave fixed effects; controls for household head
age, education, literacy, and indigenous descent; controls for the number of household members of each gender in ages brackets 0-5, 6-9, 10-12, 13-15, 16-18, and 19 or older; and controls
for the village-level marginalization index and migrant share.
Table 5: Unitary and Cooperative Test Results by Household Head Age
Panel A: All Households
Unitary test statistic
Cooperative test statistic
Panel B: Households with Younger Heads
Unitary test statistic
Cooperative test statistic
Panel C: Households with Older Heads
Unitary test statistic
Cooperative test statistic
(1)
(2)
(3)
(4)
133.23
[0.000]
20.07
[0.082]
133.43
[0.000]
20.10
[0.081]
141.71
[0.000]
20.22
[0.078]
147.66
[0.000]
20.39
[0.074]
145.88
[0.000]
19.97
[0.085]
145.97
[0.000]
20.08
[0.082]
153.84
[0.000]
19.86
[0.088]
130.65
[0.000]
17.72
[0.165]
91.22
[0.000]
12.43
[0.519]
×
91.90
[0.000]
12.28
[0.531]
90.84
[0.000]
12.87
[0.482]
×
94.37
[0.000]
13.48
[0.434]
Linear in log expenditure
Quadratic in log expenditure
×
×
Expenditure interacted with state-by-survey
×
×
wave fixed effects
Notes: Test results in column 1 are based on estimates reported in tables 3 and 4. Test results in columns 2-4 are
based on estimates from augmented regressions (not shown) that include quadratic log expenditure terms (columns
2 and 4) and interactions between the log expenditure and the state-by-survey wave fixed effects (columns 3 and 4).
These specifications follow the spirit of the linear and quadratic almost ideal demand systems discussed in section
2. Standard errors are clustered at the village level. *, **, and *** denote significance at the 10, 5, and 1% levels
respectively. The system of equations, for young and old households, is estimated simultaneously in a seeming
unrelated regression model. All equations include state-by-survey wave fixed effects; controls for household head
age, education, literacy, and indigenous descent; controls for the number of household members of each gender in
ages brackets 0-5, 6-9, 10-12, 13-15, 16-18, and 19 or older; and controls for the village-level marginalization index
and migrant share.
Table 6: Z-Conditional Cooperative Test Results by Household Head Age
Panel A: All Households
Cooperative test statistic
Panel B: Households with Younger Heads
Cooperative test statistic
Panel C: Households with Older Heads
Cooperative test statistic
Panel D: Comparing Age-Specific Curves
Test statistic for equal slopes over subsamples
(1)
(2)
(3)
(4)
52.2
[0.000]
59.4
[0.000]
55.3
[0.000]
66.4
[0.000]
57.2
[0.000]
60.0
[0.00]
61.8
[0.00]
47.6
[0.00]
14.6
[0.265]
18.3
[0.107]
15.0
[0.242]
17.2
[0.144]
31.3
[0.002]
×
20.9
[0.052]
31.8
[0.001]
×
18.0
[0.114]
Linear in log expenditure
Quadratic in log expenditure
×
×
Expenditure interacted with state-by-survey
×
×
wave fixed effects
Notes: Test results in columns 1 and 3 are from estimating models 5 and 6 respectively. Test results in columns
2 and 4 include interactions between the log consumption terms and state-by-survey wave fixed effects. Panel D
reports the results of testing the hypothesis that the coefficients on the included distribution factor, PROGRESA
eligibility, are equal for older and younger households across all budget shares. Standard errors are clustered
at the village level. p-values are shown in brackets. The system of equations, for young and old households, is
estimated simultaneously in a seeming unrelated regression model. All equations include state-by-survey wave
fixed effects; controls for household head age, education literacy, and indigenous descent; controls for the number
of household members of each gender in ages brackets 0-5, 6-9, 10-12, 13-15, 16-18, and 19 or older; and controls
for the village-level marginalization index and migrant share.
27
Table 7: Unitary and Cooperative Test Results by Household Tradition
Panel A: All Households
Unitary test statistic
Cooperative test statistic
Panel B: Households with Low Tradition Scores
Unitary test statistic
Cooperative test statistic
Panel C: Households with High Tradition Scores
Unitary test statistic
Cooperative test statistic
(1)
(2)
(3)
(4)
133.23
[0.000]
20.07
[0.082]
133.43
[0.000]
20.10
[0.081]
141.71
[0.000]
20.22
[0.078]
147.66
[0.000]
20.39
[0.074]
145.14
[0.000]
25.65
[0.010]
145.65
[0.000]
26.09
[0.008]
156.09
[0.000]
25.35
[0.012]
151.03
[0.000]
25.03
[0.013]
87.93
[0.000]
11.52
[0.593]
×
88.65
[0.000]
11.54
[0.592]
87.14
[0.000]
11.86
[0.566]
×
90.64
[0.000]
12.41
[0.521]
Linear in log expenditure
Quadratic in log expenditure
×
×
Expenditure interacted with state-by-survey
×
×
wave fixed effects
Notes: Test results are generated by estimating equations 3 and 4 (columns 1 and 2) augmented with interactions
between the log expenditure terms and the state-by-survey wave fixed effects. The tradition index is defined as
the simple average of three standardized variables: the age gap between spouses, the husband’s education and the
husband’s age. Standard errors are clustered at the village level. p-values are shown in brackets. The system
of equations, for young and old households, is estimated simultaneously in a seeming unrelated regression model.
All equations include state-by-survey wave fixed effects; controls for household head age, education literacy, and
indigenous descent; controls for the number of household members of each gender in ages brackets 0-5, 6-9, 10-12,
13-15, 16-18, and 19 or older; and controls for the village-level marginalization index and migrant share.
Table 8: Z-Conditional Cooperative Test Results by Household Tradition
Panel A: All Households
Cooperative test statistic
Panel B: Households with Low Tradition Scores
Cooperative test statistic
Panel C: Households with High Tradition Scores
Cooperative test statistic
Panel D: Comparing Subsample-Specific Curves
Test statistic for equal slopes over subsamples
(1)
(2)
(3)
(4)
52.2
[0.000]
59.4
[0.000]
55.3
[0.000]
66.4
[0.000]
63.2
[0.000]
58.3
[0.000]
68.3
[0.000]
45.0
[0.000]
19.1
[0.085]
23.2
[0.026]
19.4
[0.080]
23.2
[0.026]
36.8
[0.000]
×
24.0
[0.020]
37.2
[0.000]
×
17.7
[0.126]
Linear in log expenditure
Quadratic in log expenditure
×
×
Expenditure interacted with state-by-survey
×
×
wave fixed effects
Notes: Test results in columns 1 and 3 are from estimating models 5 and 6 respectively. Test results in columns
2 and 4 include interactions between the log consumption terms and state-by-survey wave fixed effects. Panel D
reports the results of testing the hypothesis that the coefficients on the included distribution factor, PROGRESA
eligibility, are equal for more and less traditional households across all budget shares. The tradition index is defined
as the simple average of three standardized variables: the age gap between spouses, the husband’s education and
the husband’s age. Standard errors are clustered at the village level. p-values are shown in brackets. The system of
equations, for more and less traditional households, is estimated simultaneously in a seeming unrelated regression
model. All equations include state-by-survey wave fixed effects; controls for household head age, education literacy,
and indigenous descent; controls for the number of household members of each gender in ages brackets 0-5, 6-9,
10-12, 13-15, 16-18, and 19 or older; and controls for the village-level marginalization index and migrant share.
28
Table 9: Z-Conditional Cooperative Test Results by Household Head Age in 2007
Panel A: All Households
Cooperative test statistic
Panel B: Households with Younger Household Heads
Cooperative test statistic
Panel C: Households with Older Household Heads
Cooperative test statistic
Panel D: Comparing Subsample-Specific Curves
Test statistic for equal slopes over subsamples
(1)
(2)
(3)
(4)
69.1
[0.000]
53.9
[0.000]
68.5
[0.000]
53.7
[0.000]
39.2
[0.000]
30.1
[0.003]
38.9
[0.002]
25.1
[0.015]
31.0
[0.002]
21.3
[0.047]
31.1
[0.002]
22.5
[0.032]
18.6
[0.100]
×
14.6
[0.264]
19.3
[0.082]
×
15.2
[0.230]
Linear in log expenditure
Quadratic in log expenditure
×
×
Expenditure interacted with state-by-survey
×
×
wave fixed effects
Notes: Test results in columns 1 and 3 are from estimating models 5 and 6 respectively. Test results in columns
2 and 4 include interactions between the log consumption terms and state-by-survey wave fixed effects. Panel D
reports the results of testing the hypothesis that the coefficients on the included distribution factor, wife’s income
share, are equal for older and younger households across all budget shares. Standard errors are clustered at the
village level. p-values are shown in brackets. The system of equations, for young and old households, is estimated
simultaneously in a seeming unrelated regression model. All equations include state-by-survey wave fixed effects;
controls for household head age, education, literacy, and indigenous descent; controls for the number of household
members of each gender in ages brackets 0-5, 6-9, 10-12, 13-15, 16-18, and 19 or older; and controls for the villagelevel marginalization index and migrant share.
Table 10: Z-Conditional Cooperative Test Results by Spousal Age Gap in 2007
Panel A: All Households
Cooperative test statistic
Panel B: Households with Lower Spousal Age Gaps
Cooperative test statistic
Panel C: Households with Higher Spousal Age Gaps
Cooperative test statistic
Panel D: Comparing Subsample-Specific Curves
Test statistic for equal slopes over subsamples
(1)
(2)
(3)
(4)
69.1
[0.000]
53.9
[0.000]
68.5
[0.000]
53.7
[0.000]
66.6
[0.000]
58.7
[0.000]
64.6
[0.000]
57.6
[0.000]
37.3
[0.000]
19.0
[0.088]
36.8
[0.000]
18.4
[0.105]
39.4
[0.000]
×
22.4
[0.033]
39.0
[0.000]
×
23.5
[0.024]
Linear in log expenditure
Quadratic in log expenditure
×
×
Expenditure interacted with state-by-survey
×
×
wave fixed effects
Notes: Test results in columns 1 and 3 are from estimating models 5 and 6 respectively. Test results in columns
2 and 4 include interactions between the log consumption terms and state-by-survey wave fixed effects. Panel D
reports the results of testing the hypothesis that the coefficients on the included distribution factor, wife’s income
share, are equal for older and younger households across all budget shares. Standard errors are clustered at the
village level. p-values are shown in brackets. The system of equations, for young and old households, is estimated
simultaneously in a seeming unrelated regression model. All equations include state-by-survey wave fixed effects;
controls for household head age, education, literacy, and indigenous descent; controls for the number of household
members of each gender in ages brackets 0-5, 6-9, 10-12, 13-15, 16-18, and 19 or older; and controls for the villagelevel marginalization index and migrant share.
29
30
Difference from baseline
(old-young, control; α3 )
Difference from baseline
(old-young, treatment; α3 + α6 )
Treatment effect
(young; α4 )
Difference in treatment
effect (old-young; α4 + α7 )
Baseline mean
(st. dev.)
Drop households with
potentially eligible kids
Controls for
wealth and education
N
Completed
primary school
(2)
0.047
[0.018]**
0.031
[0.013]**
-0.022
[0.020]
0.027
[0.023]
0.56
(0.38)
Yes
Yes
12,587
Completed
primary school
(1)
0.038
[0.019]**
0.016
[0.013]
-0.023
[0.020]
0.030
[0.023]
0.56
(0.38)
Yes
No
12,642
No
19,834
Completed
primary school
(3)
0.049
[0.015]***
0.037
[0.011]***
-0.019
[0.017]
0.033
[0.018]*
0.51
(0.42)
No
No
12,738
Completed
secondary school
(4)
0.020
[0.007]***
0.026
[0.007]***
0.002
[0.013]
-0.008
[0.017]
0.07
(0.12)
Yes
Yes
12,682
Completed
secondary school
(5)
0.013
[0.008]*
0.022
[0.007]***
0.003
[0.013]
-0.009
[0.017]
0.07
(0.12)
Yes
No
20,472
Completed
secondary school
(6)
0.033
[0.009]***
0.035
[0.006]***
0.009
[0.011]
-0.008
[0.015]
0.16
(0.26)
No
No
12,623
Enrolled and
aged 6-11
(7)
-0.005
[0.010]
-0.029
[0.007]***
-0.008
[0.010]
0.037
[0.014]***
0.95
(0.18)
Yes
Yes
12,574
Enrolled and
aged 6-11
(8)
-0.003
[0.010]
-0.026
[0.007]***
-0.008
[0.010]
0.038
[0.014]***
0.95
(0.18)
Yes
No
28,018
Enrolled and
aged 6-11
(9)
-0.008
[0.011]
-0.014
[0.006]**
-0.003
[0.010]
0.021
[0.012]*
0.92
(0.24)
No
No
15,306
Enrolled and
aged 12-16
(10)
-0.042
[0.020]**
-0.064
[0.017]***
0.047
[0.023]**
0.031
[0.026]
0.55
(0.42)
Yes
Yes
15,253
Enrolled and
and aged 12-16
(11)
-0.004
[0.020]
-0.034
[0.017]**
0.046
[0.023]**
0.037
[0.026]
0.55
(0.42)
Yes
Table 11: Educational Attainment: share of household offspring (a) who completed primary school (for those aged 11-16) and secondary school (for
all ages) and (b) aged 6-11 or 12-16 and currently enrolled in school.
No
22,934
Enrolled and
and aged 12-16
(12)
0.007
[0.018]
-0.009
[0.014]
0.053
[0.019]***
0.007
[0.021]
0.58
(0.43)
No